Infographic-Statistic-to-Spreadsheet-Cell Audit Prompt

This prompt is designed for a data analyst or verification engineer who needs to reconcile every statistic displayed in an infographic against the exact spreadsheet cell that supposedly generated it. The job-to-be-done is a cell-level audit trail, not a general summary or a high-level accuracy opinion. The ideal user already has a structured description of the infographic's visual elements—chart types, axis ranges, data labels, color scales—and a structured representation of the source spreadsheet data. The prompt assumes these two inputs are already extracted and formatted; it does not perform OCR, chart parsing, or spreadsheet ingestion. Use this prompt when the cost of a misrepresented statistic is high: regulatory filings, investor reports, published research, or public-facing dashboards where a truncated axis or a misleading color scale can cause real harm.

The prompt produces a reconciliation table that maps each infographic statistic to its spreadsheet origin, flags discrepancies, identifies rounding errors, and detects visual-encoding misrepresentations. It checks for truncated axes that exaggerate differences, color scales that imply false thresholds, and aggregation-scope mismatches where a chart shows a subset but the label implies a total. The output is structured for downstream validation: each row includes the infographic value, the spreadsheet cell reference, the expected value, the discrepancy type, and a severity flag. This structure lets you pipe results directly into a review queue or a programmatic validator. The prompt also includes explicit instructions to abstain when the mapping is ambiguous rather than guessing, which is critical for audit workflows where a false match is worse than a flagged unknown.

Do not use this prompt for general chart-reading tasks, for generating infographics from spreadsheets, or for verifying claims that span multiple source documents without a single canonical spreadsheet. It is not a replacement for a statistical analysis prompt—it does not interpret trends, assess significance, or comment on data quality beyond the reconciliation scope. If your infographic description was generated by a vision model, validate that description for hallucinations before feeding it into this prompt; a hallucinated axis label will produce a false discrepancy cascade. After running this prompt, always route severity-flagged rows to a human reviewer and log the full reconciliation table as an audit artifact. The next step is to wire this prompt into a verification pipeline that includes input validation, structured output parsing, and a human-in-the-loop escalation path for high-severity discrepancies.

This prompt template is designed to be pasted directly into your AI system for the specific task of auditing statistics displayed in an infographic against the raw data in a source spreadsheet. It is built for data analysts and verification engineers who need to move beyond visual inspection and produce a structured, cell-level reconciliation. The template forces the model to map visual encodings to numerical values, check for rounding errors, and flag aggregation mismatches, turning a manual audit into a repeatable, traceable process.

textCopy
You are an audit assistant specialized in reconciling data visualizations with source spreadsheets. Your task is to compare the statistics presented in an infographic against the raw data in a provided spreadsheet and produce a structured audit report.

## Input
- Infographic Description: [INFOGRAPHIC_DESCRIPTION]
- Source Spreadsheet Data: [SPREADSHEET_DATA]

## Audit Instructions
For every statistic you can identify in the infographic description, perform the following steps:
1.  **Value Extraction:** State the exact numerical value and label as presented in the infographic.
2.  **Visual Encoding Mapping:** Describe how the value is visually represented (e.g., bar height, pie slice angle, color intensity, explicit text label). If the value is derived from a visual encoding, explain your calculation.
3.  **Source Data Location:** Identify the precise cell range, column, or pivot table field in the [SPREADSHEET_DATA] that should contain the source value. If no matching source data is found, flag it as "Source Not Found."
4.  **Reconciliation:** Compare the infographic value to the source data value. Calculate the absolute and percentage difference. Apply a default tolerance of ±0.5 for rounding.
5.  **Discrepancy Flag:** If the difference exceeds the tolerance, flag the item as a "Discrepancy" and propose a root cause from this list: [Rounding Error, Aggregation Scope Mismatch, Incorrect Source Reference, Visual Encoding Misinterpretation, Data Error, Other].
6.  **Audit Note:** Add any clarifying context, such as unit conversions applied or assumptions made.

## Output Format
Return a single JSON object with the following schema:
{
  "audit_report": [
    {
      "infographic_label": "string",
      "infographic_value": "number",
      "visual_encoding_notes": "string",
      "source_data_location": "string",
      "source_data_value": "number or null",
      "absolute_difference": "number or null",
      "percentage_difference": "number or null",
      "discrepancy_flag": "boolean",
      "proposed_root_cause": "string or null",
      "audit_note": "string"
    }
  ],
  "summary": {
    "total_statistics_audited": "number",
    "total_discrepancies_found": "number",
    "critical_findings": ["string"]
  }
}

## Constraints
- Do not invent data. If a value cannot be determined from the infographic description or source data, use `null`.
- Explicitly state any assumptions about aggregation (e.g., sum, average) or units.
- Flag any instance where a visual encoding (like a color scale) is ambiguous and could represent multiple values.

After pasting this template, the most critical adaptation is populating the [INFOGRAPHIC_DESCRIPTION] and [SPREADSHEET_DATA] placeholders with high-quality, structured inputs. For the infographic, a purely textual description of a chart is often insufficient; you should include a structured breakdown of axes, labels, legend items, and data point coordinates. For the spreadsheet, providing the data as a CSV or markdown table string is far more reliable than a screenshot. The next step is to integrate this prompt into a validation harness that checks the output JSON against the defined schema before the results are accepted, preventing malformed reports from entering your audit trail.

This prompt is designed to operate as a single step inside a larger audit pipeline. The typical harness receives an infographic image and a source spreadsheet as input, runs the prompt, and then validates the structured output before any downstream action. Because the prompt performs cell-level reconciliation, the harness must enforce strict output schema validation, handle model refusals or malformed JSON, and route ambiguous findings to a human reviewer. Do not treat this prompt as a standalone one-shot tool; it should be wrapped in application logic that controls retries, logs every audit run, and preserves the original infographic and spreadsheet for traceability.

The implementation should follow a validate → retry → escalate pattern. After the model returns a JSON response, validate it against the expected schema: every claim object must have a non-null infographic_value, spreadsheet_cell_reference, spreadsheet_value, match_status (one of exact_match, within_tolerance, mismatch, not_found_in_spreadsheet, not_found_in_infographic), and a discrepancy_detail string. If validation fails or the model returns a FALLBACK marker, retry once with a simplified version of the prompt that asks only for the mismatches. If the retry also fails, log the raw response and route the entire infographic-spreadsheet pair to a human review queue with the partial results attached. For high-stakes audits, always require human sign-off on any mismatch or not_found finding before the output is considered final.

Model choice matters here. Use a model with strong vision and structured-output capabilities, such as GPT-4o or Claude 3.5 Sonnet, and set response_format to json_object (or use the model's equivalent strict JSON mode) with the output schema provided in the prompt. For batch processing across many infographics, implement rate limiting, cost tracking per audit run, and a dead-letter queue for images that exceed token limits or trigger content filters. Store every audit result—including the prompt version, model version, timestamp, raw response, and human review decision—in an append-only audit log. This log becomes the evidence trail if any finding is later disputed. Finally, test the harness end-to-end with a golden dataset of infographic-spreadsheet pairs where you know the expected discrepancies, including edge cases like color-encoded values, truncated axis labels, and aggregation-scope mismatches.